Digital Twin (DT) technology has emerged as a transformative solution for real-time monitoring, fault diagnosis, and predictive maintenance in power electronic converters. Recent studies have shown that, as a black-box modeling approach, neural network (NN) based models of power electronics converter can offer some advantages over physical models. Leveraged by this, this paper explores feasibility of NN based DT for a three-phase inverter. A physics-informed NN is proposed to represent the conventional power switch-based physical model of the three-phase inverter. Experimental results demonstrate that the NN-based DT achieves high accuracy in predicting the dynamic behavior of the physical inverter while significantly reducing computational time. This indicates that the NN-based DT can be deployed on lower-end microcomputers, enhancing reactivity and performance in real-time applications.

This paper presents a magnetic coupler design method for wireless power transfer (WPT) systems, which is based on machine-learning algorithms. A synthetic dataset generated by ANSYS-Maxwell is used for training and evaluating machine-learning models. The trained model can obtain the optimal values of the coil inner radius Riand number of turns N, when other coil design parameters such as the coil outer radius Ro, the wire diameter Rw, and the coil inductance L, are given based on the application environment. The proposed method provides a practical solution to design of magnetic couplers with ferrite cores and also accelerates the design compared with conventional methods.

Nowadays, the data-driven approaches to modeling power electronic systems are mostly based on common neural network (NN) structures without any prior knowledge of physics. However, these approaches often demand extensive data across a wide range of inputs and may encounter overfitting issues. To address this issue, this paper proposes a new data-driven modeling approach for DC-DC converters based on neural ordinary differential equations (ODE). This approach incorporates prior knowledge about numerical ODE solvers into the NN structures, leading to a substantial improvement in generalization performance. In this method, the kernel function of numerical integration is replaced by a fully-connected NN, and the system ODE is solved through common numerical solvers iteratively in time. The results of our study demonstrate a significant enhancement in generalization performance across various input frequencies using variable step solvers while maintaining excellent precision.

Nowadays, conducting detailed numerical simulation of power electronic (PE) converters is time-consuming due to the iterative solution of nonlinear equations. While simplified numerical models can reduce the computational burden, they are unable to guarantee fidelity in highly nonlinear systems. To accelerate the detailed numerical simulation while maintaining high fidelity, this paper proposes a novel data-driven approach to modeling PE converters based on neural ordinary differential equations (ODE). This model eliminates the iterative solution of nonlinear equations by being data-driven and thus accelerates the simulation. In addition, to enhance the model fidelity, this approach incorporates prior knowledge about numerical ODE solvers into the model structures and distinguishes multi-scale characteristics from the dataset using a specific filtered MSE loss function. Experiment results demonstrate significant improvement of calculation speed and reduction of computational overheads compared to the detailed numerical methods. In addition, the proposed model has shown excellent modeling fidelity across various input frequencies using variable step solvers.